The neutrality commitment

Draft. This document has not been adopted and is not in force. It is published here so it can be read and reviewed before it is.

Also: Terms · Privacy

Status: DRAFT. Not adopted. Drafted 2026-09-17 for MODEL-70. Version neutrality-v1, which is the same string the machine-readable copy carries.

The rule

Charging the consumer of a recommendation is compatible with being an honest broker. Charging the subjects of one is not.

An agent paying to ask "what should I use" has no stake in which answer comes back. It wants the true one, which is exactly what it is paying for. A model provider paying to appear is buying placement, and every ranking afterwards is suspect whether or not it was actually influenced. The second is not a conflict to be managed with a disclosure line. It is a different product, and we are not in it.

The commitment

No referral fees, no paid placement, no provider-paid visibility, permanently.

Stated as the things we are not able to do rather than the things we promise not to do, because the first kind can be checked:

• We take no money from the subjects of a ranking. No provider, host, gateway or vendor can buy inclusion, position, prominence or a mention. There is no rate card, because there is no price.
• We do not proxy inference tokens. We recommend and hand off. Your inference traffic goes to the provider, gateway or local runtime directly. We therefore earn nothing that scales with your token spend, and have no reason to prefer the route that spends more.
• We do not hold your prompts. A request carries a locally computed profile, not prompt text, so there is nothing to sell, leak or mine. See the privacy statement.
• We are neutral in sourcing, not only in money. No provider, gateway, hosting route or inference vendor is privileged at any stage: not in the ranking, not in the tie-breaks, not in a default hosting suggestion, not in routing advice, and not by any commercial relationship of the operator's. Refusing payment from the subjects of a ranking is the easy half. The harder half is refusing the version where nobody pays and the advice leans anyway.

"Permanently" is the load-bearing word. A commitment that lasts until the offer is good enough is a price, not a commitment.

How to check it, without trusting us

The commitment is published as data, next to the ranking floors, in the same object that describes the policy an answer was computed under:

GET https://modelspec.dev/api/rank/profiles.json

No key. No account. No rate limit. It is a static file on the same export the sites and the CLI read.

.ranking_policy.neutrality

{
  "version": "neutrality-v1",
  "operator": "Sparks & Sawdust LLC",
  "rule": "Charging the consumer of a recommendation is compatible with being an honest broker. Charging the subjects of one is not.",
  "pledge": "No referral fees, no paid placement, no provider-paid visibility, permanently.",
  "permanent": true,
  "assertions": {
    "accepts_referral_fees": false,
    "accepts_paid_placement": false,
    "accepts_provider_paid_visibility": false,
    "proxies_inference_tokens": false,
    "stores_customer_prompts": false
  },
  "source_neutral_at": ["ranking", "tie_breaks", "hosting_suggestions", "route_advice"],
  "charges": "the consumer of a recommendation, never its subjects"
}

The same block appears in the policy object of every response from POST https://api.modelspec.dev/v1/rank, so an agent reads the commitment out of the very answer it acted on rather than out of a page it never fetched.

Three further things are checkable without asking us:

• The method is public. api/ranking/engine.py is the scorer, and the floors it publishes (min_benchmark_coverage, min_benchmark_count) are the ones it applies. The rank endpoint runs that file, vendored verbatim, and the test suite fails if the two ever disagree.
• The evidence is dated. Every card cites its sources with the date each was read, and each ranked row reports its evidence_basis.
• Absence is reported, not hidden. Models without enough evidence are returned as unranked rather than quietly ranked last, and the response says how many there were. A ranking that hid them would look better and say less.

If we ever break this commitment, the JSON breaks first. That is the point of putting it there.